1. Introduction
Gāndhārī, a Middle Indo-Aryan language spoken between the third century bce and the fifth century ce and written in the Kharoṣṭhī script, has come to us in the form of inscriptions on stone or metal (Salomon Reference Salomon1998: 74–9), on coins (Salomon Reference Salomon1998: 209–18), as well as a range of manuscripts written on wood (Boyer, Rapson, and Senart Reference Boyer, Rapson and Senart1920; Burrow Reference Burrow1937, Reference Burrow1940), birch bark, or palm leaf (Brough Reference Brough1962).
Spoken in an area at the crossroads of several other cultures, Gāndhārī came to play an important role in the region as an administrative and religious language, and neighbouring languages came to borrow Gāndhārī words, such as Tocharian B anās/Tocharian A ānās “miserable” (Gdh. anasa < Skt. anātha “without protector”, showing the regular development th > s/V _ V) (Carling and Pinault Reference Carling and Georges-Jean2023) or Chinese 曇摩 /dəm.mɑ/ (Gdh. dhaṃma) (Coblin Reference Coblin1983: 248).
Beyond its administrative role, Gāndhārī is more importantly the language of the oldest extant Buddhist texts in any language (Salomon Reference Salomon1999: 9), and the language is also thought to be the language from which were translated the first Chinese Buddhist texts in the second half of the second century ce, by the translation teams around Ān Shìgāo 安世高 (fl. 148–170) (Zacchetti Reference Zacchetti2019: 630) and Lokakṣema (Zhī Lóujiāchèn 支婁迦讖, fl. 147–189) (Harrison Reference Harrison2019: 700). The evidence for Gāndhārī being the source language of these earliest Chinese Buddhist texts is the use of phonetic transcriptions for technical Buddhist terms (Baley, Hill, and Caldwell Reference Baley, Hill and Caldwell2023) that are better explained if the phonology of Gāndhārī is assumed.Footnote 1
Because of its central role in the study of early Buddhism – both as the language of the earliest extant Buddhist texts and as the source language of the earliest Chinese Buddhist translations – the study of Gāndhārī’s phonology – both synchronic and diachronic – plays an essential role in helping us understand these early texts and their transmission. Since the early forays into Gāndhārī studies, a large body of academic literature has dealt with its phonology and its development from Sanskrit, and a good introduction to a corpus-wide analysis can be found in Baums (Reference Baums2019).
One topic that has so far not received sufficient dedicated attention is the development of Sanskrit velar stops: beyond the uncontroversial treatment of aspirated stops (cf. for instance Brough Reference Brough1962: 93, or Baums Reference Baums2009: 143), the development of unaspirated velars is less well understood. This is in part due to the deficient nature of the Kharoṣṭhī script, but also because of the nature of the corpus: not only do different documents represent different periods of the development of the language (Fussman Reference Fussman1989: 455–65), but there is also a layering of Gāndhārī phonology with more Sanskritized elements (Salomon Reference Salomon2001: 245), resulting in individual manuscripts sometimes appearing inconsistent.Footnote 2 So far, the field has not reached a consensus, some considering that unaspirated velar stops have evolved into a velar fricative (usually the voiced [ɣ] because it is spelled with a modified <g>, transcribed g̱) (Burrow Reference Burrow1937: 6) while others think this spelling only reflected scribal idiosyncrasies, the velars supposedly having entirely disappeared from the language or remaining as a palatal approximant [j] (whether written <y> or absent from the spelling) (Baums Reference Baums2009: 140).
In the rest of this article, we propose to solve this question of the development of Sanskrit velars into Gāndhārī. For completeness, we also include the uncontroversial development of aspirated velar stops. The structure of the article is as follows: first, we introduce our methodology, the decisions we have made, and the tools we have used. Then, we focus on the development of velars in different environments: word-initial, word-final, on either side of another consonant, and finally in intervocalic position. Since most of the difficulty lies in the intervocalic context, this section is itself divided into the analysis of aspirated stops on the one hand, and a longer analysis of the unaspirated ones on the other. In the process, we show that the development of unaspirated velar stops was clear and that the spelling practices found throughout the corpus reflect actual phonetic phenomena; the corpus itself is analysed in a diachronic fashion,Footnote 3 giving hints into the dating of the sound changes pertaining to velars. Finally, we summarize our discoveries and provide further phonetic evidence in other languages.
2. Methodology
Below, we briefly present the approach we have taken to build a dataset, annotate it, and analyse it.
2.1. Corpus dataset
Our study uses the online Gāndhārī dictionary found at Baums and Glass (Reference Baums and Glass2002). Baums and Glass rely on previously published text editions of theirs and others, and for each attested word form found in a Gāndhārī document (inscription, manuscript, coins, tablets, etc.), they list the position in which it is found in the document, the corresponding standardized Gāndhārī lemma, nominal or verbal inflection information (number, gender, case, etc.), and the corresponding Sanskrit and Pali words.Footnote 4 See Figure 1 for an example.
Figure 1. Entry in Baums and Glass (Reference Baums and Glass2002) for the word “dhaṃma”.
The data can be queried either through the standardized lemma, the specific attested form, the Sanskrit equivalent, or the English translation. For each of these modalities, regular expressions can be used, so that one can for instance search for all words having a <k> in Sanskrit and study their Gāndhārī outcomes.
As the focus of our study is the development of Sanskrit velar stops, we have retrieved all entries where a Sanskrit equivalent has been identified and where that Sanskrit equivalent contains a k or g, which naturally includes the aspirated series with kh and gh. These data have then been arranged in a table, with attested word form per line, along with its standardized Gāndhārī lemma, Sanskrit equivalent, and document of origin. This dataset contains some 5,000 entries.
2.2. Dataset annotation
To analyse the development of Sanskrit k, g, kh, and gh across the corpus, we annotate the outcome of that velar in each attestation. For instance, Skt. dāraka- “child” is attested in AnavL as daraghasa and in AvL6 as daraga, which we annotate respectively as gh and g, while Skt. adhikam is attested in SC2 as as̱io, for which we annotate the outcome of the k as ∅ (i.e. an absence of character). The set of possible outcomes for the velar stops in our annotation scheme is {k, ḵ, kh, g, g̱, gh, h, y, ∅}, where ḵ and g̱ transcribe the Kharoṣṭhī characters for k and g, to which the scribe had added one of a variety of foot marks, as discussed in Glass (Reference Glass2000: 56) and shown in Figure 2.
Figure 2. The footmark types found (amongst other characters) under g to denote g̱. Image taken from Glass’s classification (Glass Reference Glass2000: 56).
2.3. Presentation of the results
With the whole dataset annotated, we can then turn to analysing trends in the outcome of the velar stops in specific environments. In our tables, we use the document code names as found at gandhari.org, such as AvL1 for the Avadānas (BL 1.2). As the whole corpus available at gandhari.org contains dozens of documents, many of which are very short and might only contain a very small number of words of interest (i.e. derived from a Sanskrit ancestor having a velar stop), we choose to restrict our presentation of the outcome counts to the 30 documents containing the most words of interest. In the process, we also decided to group together the various families of documents indicated with a number suffix: for instance, for the Avadānas texts found in BL1.2, BL2, BL3A, BL4.2, BL12+14.2, BL16+25.2, BL21.2 and listed with the codes AvL1 through AvL7, we group them under the heading AvL. We apply this grouping approach to all documents having a number suffix,Footnote 5 except for the items using CKM (manuscripts), CKI (inscriptions), and CKD (documents) for which the code represents an otherwise unclassified item. We have found this approach to lead to data that is more legible and, thanks to the coherence of the groupings, better trends can be highlighted.
3. Non-intervocalic velars
Velar stops that are not in intervocalic position behave much more regularly than intervocalic ones and are generally uncontroversial. For completeness, we treat them first in the four possible contexts: word-initial, word-final, following or preceding a consonant.
3.1. Word-initial velars
In word-initial position, the four velar stops are nearly always preserved across the whole corpus, as has been widely noted. There are two main exceptions: first, the so-called “scribe 1” of the EĀL (Allon Reference Allon2001: 68), Dhp-GL (Lenz Reference Lenz2003: 41), and AnavL (Salomon Reference Salomon2008: 96) who never writes g in any context and instead writes gh. We consider this an idiosyncrasy and not indicative of a phonological difference, as seems to be implied by the three authors dealing with this scribe’s texts above.
Second exception, Sanskrit word-initial gh is predominantly found as g in Gāndhārī, except in words inherited from Skt. ghr- or ghṛ-, which preserved the gh- in most attestations. This suggests a regular de-aspiration of the voiced aspirated velar stop, at least in initial position, that would have happened before the earliest extant Gāndhārī texts.Footnote 6
Beyond the regular pattern of preservation of k-, kh-, and g-, the de-aspiration of gh-, and the scribal idiosyncrasies, we find a very small number of exceptions:
• Skt. kula “family, herd” is found as khulana in a few texts which otherwise show initial k reflexes.
• kim loses its initial velar and appears as iṃ in several texts, always in the pair “ma iṃ” or “na iṃ”. This likely occurred because the word is an enclitic – the k behaves as if it were in intervocalic position.
• A single instance of Skt. Kuṣāṇa-vaṃśa-saṃvardhaka > Gdh. guṣaṇavaśasaṃvardhaka in CKI 149.
• Initial k appears as ḵ in four attestations in the SĀS documents: Skt. kṛta > ḵiḏa (SĀS5), Skt. karaṇīya > ḵaraṇia and Skt. kim > gen.sg. ḵis̱a (SĀS6), and Skt. karaṇīya > ḵaraṇao (SĀS7).
• Skt. khalu “indeed” is commonly eroded to shorter forms, some of which do not have kh: ho, kho, o, khu, hu, gu, kha, u, ku, khalu. This is likely due to khalu being an unstressed enclitic particle.
These exceptions are so few that they should be of no concern to us regarding the regularity of the rules established above, and summarized below:
k > k / # _
kʰ > kʰ / # _
g > g / # _
\begin{equation*}
\text{g}^{\text{h}} \; \gt \;
\begin{cases}
\text{g}^\text{h} / \# \_ r \\
\text{g} / \# \_
\end{cases}\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\ \
\end{equation*}3.2. Word-final velars
To our knowledge, word-final velar stops have not been discussed previously, probably due to their rarity. Of the four velar stops, only -k is found in Sanskrit, in a handful of adverbs and connecting words: samyak “properly”, yādṛk “like”, tiryak “across”, dhik “fie (interjection)”, and viṣvak “in all directions”. In all of these, k disappears from the spelling in Gāndhārī, and we assume that this reflects the phonetic reality of its disappearance:
k > ∅ / _ #
3.3. Velars following other consonants
As the conventional wisdom in studies of Gāndhārī phonology is that most consonants lenite in intervocalic position (Baums Reference Baums2019: 4), it is interesting to study the cases where the consonant is not in intervocalic position, so as to establish a baseline for the development of individual consonants. In this section, we study the outcome of velar stops preceded by a consonant.
k is found after the following consonants: l, r, s, t, ṃ, ṅ, ṣ, and ṛ:Footnote 7
• In rk, the cluster is preserved, although the spelling might suggest sporadic metathesis of the r. In the cluster tk, the t is dropped and k is preserved; this has usually been interpreted (e.g. Allon Reference Allon2001: 89) as a case of regressive assimilation resulting in the gemination of k. Assuming this assimilation occurred, whether gemination was retained in Gāndhārī by the time of our corpus is uncertain. In lk, the preceding consonant is dropped and k is preserved, possibly also hinting at a geminated k, as found in other Middle Indo-Aryan languages.
• The cluster sk regularly yields kh (with s dropped), with the exception of Skt. skandha “factor” and its compounds, which all yield k.
• The cluster ṣk similarly drops the s, but the outcome is split equally between k and kh, and some documents show both outcomes. We have currently no explanation for this fluctuation, but perhaps dating or dialectal information could help understand this behaviour.
• ṃk yields g in all texts, and while the nasal is usually not written, Glass (Reference Glass2007: 106) comments that the lack of lenition is evidence of its preservation. This can be seen in the texts that normally preserve k (DhpK) or yield y (SĀS) in intervocalic position, as these texts also show ṃk > g, which confirms that the nasal must have been preserved and caused the voicing of the velar stop.
• ṅk yields g in all texts. As with ṃk, we take this as evidence that the nasal was preserved and caused the voicing.
• There are few examples of cluster Skt. ṛk, whose outcome is similar to k in intervocalic position and suggest that the sound change ṛ > ri happened before the various changes happening to velars in intervocalic position and described further below.
g is found after the following consonants: d, l, r, ḍ, ṃ, ṅ, ṛ:
• As for k clusters, rg is preserved as a cluster. As for tk above, dg yielded g, interpreted as a geminate (Salomon Reference Salomon2000: 88). The outcome of ḍg as rg suggests that the development ḍ > ɽ – i.e. of a stop into a flap – happened earlier than the time the stop+velar clusters (tk, ḍg) assimilated to their stop.Footnote 8 The outcome of lg is g, perhaps also hinting at a geminated g.
• The clusters ṃg and ṅg, like ṃk and ṅk, preserve the g even in documents that otherwise show a lenition of intervocalic g to /j/ or ∅, bringing further evidence that – regardless of it being represented in spelling or not – the preceding nasal was preserved.
• The cluster ṛg is mainly found in the Skt. word mṛga “deer” and – like ṛk – also yields outcomes that suggest an early development ṛ > ri.
kh is found after the following consonants: r, s, ḥ, ṃ, ṅ. As opposed to k and kh behaving very differently in intervocalic position as will be seen later, kh behaves similarly to k when following a consonant:
• The cluster ḥkh yields kh, the r of rkh is preserved as in the development of rk and rg.
• The Skt. cluster skh loses its s and yields kh, like sk. The complete absence of any attested sk or skh in Gāndhārī tells us this change must have happened very early, but precludes any certainty regarding the development of these individual clusters. In the case of sk, we follow the suggestion of a reviewer of an early draft of this article, and propose the sequence sk > ʰk > kʰ.
• ṃkh and ṅkh both yield gh, as ṃk and ṅk both yield g, adding further evidence of the retention of the nasal, at least until the stage where intervocalic unaspirated velar stops lenited.
gh is rather rare in clusters, found after the following consonants: d, r, ṃ. The development of these clusters is generally coherent with the previous sections:
• dgh drops its stop and yields gh. As for tk and dg, this has been interpreted as a sign of gemination.
• ṃgh is found in Skt. saṃgha “community” and yields gh, with the same conclusion that the nasal helped preserve the stop, since – as will be discussed later – gh in intervocalic position regularly yields h.
• rgh is found in Skt. dīrgha; all texts spell a metathesis of r as dri-, and the velar is variously spelled as g, gh, kh. We have no explanation to offer for the inconsistency in spelling of this word, found nine times throughout the whole corpus.
The changes can be summarized as such:
kʰ > gʱ / [ +nasal ] _
k > g / [ +nasal ] _
skʰ > kʰ
sk > ʰk > kʰ
3.4. Velars preceding other consonants
A large part of the words showing a velar stop followed by another consonant are compound words, the velar stop being the last phoneme of the first part of the compound, and the other consonant the initial phoneme of the second part. In such cases, things behave as they usually do in compounds, i.e. the velar stop behaves as if it were at the end of a word and disappears. In the paragraphs below, we leave such cases out of the discussion. Overall, the scarcity of the material means that our present analysis should be considered tentative.
k is found before the following consonants: l, r, t, v, ṛ, and ṣ.
• kl is found in two words and their compounds: kleśa “defilement” always yields kileśa, while śukla “bright/light” always leads śukra. No specific behaviour can be inferred from these two sole examples.
• kv preserves its v in initial position and loses it otherwise.
• kt usually loses its k, interpreted as a gemination of t, although some kt are still found in spelling.
• kr is preserved very consistently.
• kṛ has two different outcomes, kri and ki. The outcome seems to be decided on a document basis; we can exclude the possible inability of some scribes to spell a kr cluster as the documents writing ki do have other kr clusters (cf. previous item). This leaves us with the tentative explanation that dialectal differences are at stake. If that is the case, and since kr itself is consistently preserved throughout the corpus, the route kṛ > kri > ki has to be excluded. Instead, since ṛ occasionally yields ir instead of ri, we propose kṛ > kir > ki/_ C, on the ground that r > ∅/_ C is a regular sound law in Gāndhārī (Baums Reference Baums2019: 6).
• Skt. kṣ usually yields Gdh. kṣ, which is written with a specific ligature in Gāndhārī whose transcription as kṣ has been attributed on the ground of its usual correspondence with Sanskrit kṣ. The actual pronunciation of this ligature is still debated, the debate residing in whether the cluster is pronounced with an initial /ʈ/ (Bailey Reference Bailey1946: 774) or initial /k/ (Schoubben Reference Schoubben2022: 150), a debate we do not aim to resolve here. What is interesting, however, and may contribute to the debate, is that beyond the more common outcome kṣ, we also find kh in a variety of texts and in a variety of environments,Footnote 9 a phenomenon we cannot yet explain but which would tend to favour a /k/ initial in the phonetic realization of the <kṣ> graph, at least in some contexts.
g is found before the following consonants: n, r, ṛ.
• gn: agni “fire” and its compounds all yield g; the fact that g is attested in Mnrs (and not g̱), in BC (instead of y), and in Khvs (instead of ∅), suggests that /n/ was pronounced at least until after the lenition of intervocalic g.
• gr: similar to gn, nearly always yields either gr or g (except in a few instances of kr), and g in EĀL suggests that the /r/ persisted at least until after the time of the lenition of intervocalic g, and was in all likelihood simply not written.
• gṛ: found in the various conjugations of the verb Skt. √gr(b)h- “take”. Our interpretation of the evidence is here is similar to kṛ presented above.
The changes relevant to velar stops can be summarized as such:
k > ∅ / _ t
\begin{equation*}
\text{k}\upsilon \; \gt \;
\begin{cases}
\text{k}\upsilon / \# \_ \\
\text{k}
\end{cases}\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\quad
\end{equation*}4. Intervocalic velars
As the cases where velar stops are found with a consonant or a word boundary on either side have been covered, we now have a firm baseline to evaluate the specific development of velar stops in intervocalic position.
4.1. Velars at compound boundary
Previous literature has noted that intervocalic velar stops, and any obstruent in general, behaved differently when a compound boundary occurred on either side of the velar (Konow Reference Konow1929: xcviii), effectively blocking the lenition.Footnote 10 This makes sense both if the compound was created after the period when the intervocalic lenition sound change was operating, and if the occurrences of the morpheme in unchanged positions provided sufficient analogical pressure for its preservation in other positions.Footnote 11
In our dataset, we annotated each entry for which the velar is not word-initial or word-final (incl. word-final suffixes like -ka), marking whether a compound boundary was present before the velar. This affords us analysing the behaviour of intervocalic velars both in compound boundary environments and outside.
Overall, we reach the same conclusion as previous scholars: compound boundary blocks lenition and the velar stop is preserved, i.e. the second part of the compound behaves as if it were a separate word and the velar was in word-initial position. It is worth noting here that verbal and nominal prefixes (e.g. abhi- “towards”, adhi- “over, besides”, ā- “near, towards”, anu- “after, near”, pra- “before, forward”, prati- “towards”, upa- “towards, near to”) themselves do not have this blocking effect, and we find lenition of the velar after them.
4.2. Velars outside compound boundaries
True intervocalic velars represent the most interesting segment of our study, because they are the ones where previous scholarship has shown the most confusion. We approach this in four parts: first, we deal with the uncontroversial aspirated velars; then, we study the behaviour of words with a -ka suffix and demonstrate that the velar there shows a very regular development; then, we look at k in environments other than this -ka suffix, and finally we study the behaviour of intervocalic g in light of our understanding of the development of intervocalic k.
4.2.1. Aspirated velars kh and gh
Previous scholarship notes that Skt. kh and gh in intervocalic position always yield a spelled h (for instance, Salomon Reference Salomon2000: 82; Glass Reference Glass2007: 114; Schlosser Reference Schlosser2022: 78).
We reach the same conclusion, but it is worth noting that some documents still spell kh in the words Skt. mukha “mouth, face, countenance”, sukha “pleasant”, and in the conjugation of the root √likh “to write”. This is the case for the texts in Shāh and Mān, which generally display a very archaic phonology, but also in EĀBm, ASP and the AvL. The outcome of gh follows the same pattern and yields h throughout the corpus, with a few exceptions: Bhks writes gh in its only attestation, DhpK writes k twice, and NirdL2 writes nothing in all attestations of Skt. pratigha (while it writes h elsewhere), and so does MSūB. The relative scarcity of data precludes any strong conclusion, as what we witness might be the result of sampling.
4.2.2. The case of the -ka suffix
Brough (Reference Brough1962: 91) commented that the weakening of k was found “with fair consistency” and was “especially common in stem-final position – frequent therefore for the suffix (-ka), although not limited to it”. The particular tendency for the suffix -ka to lenite was also noted elsewhere, for instance in Salomon (Reference Salomon2000: 82) and Allon (Reference Allon2001: 80).
Yet, other scholars comment that the variation in spelling is a cause for doubt. The evolution of scholarly positions on this topic is interesting. Brough (Reference Brough1962: 86) ascribes to g̱ (then transcribed ǵ) the value of [ɣ] in some cases and [χ] in others and considers the matter “generally accepted”,Footnote 12 but also notes that – for intervocalic single stops – “such variation [in spelling] is for the most part a matter of spelling”.
Glass (Reference Glass2000: 56), in the first volume of the Gandhāran Buddhist Texts collection, comments that “these flourishes [such as the one at the bottom of g, transcribed g̱] do not have any phonetic value”, and later in the volume Glass (Reference Glass2000: 82) notes “inconsistent treatment of original -k- is typical of Gāndhārī/Kharoṣṭhī texts in general and evidently shows that it was pronounced weakly, if at all”, but also mentions “a diacritically marked variety of g, apparently represents a fricative pronunciation (/ɣ/)”.
In a further volume of the collection, Lenz (Reference Lenz2003: 118) remarks “a distinct character (transcribed ǵa), believed to have a fricative pronunciation /ɣ/ […]. However, a distinct phoneme is probably not indicated in our text”.Footnote 13 Based on this comment, Lenz makes no attempt to transcribe this distinct letter, transcribing g where other volumes of the collection write g̱, an issue we discuss in Appendix B3.Footnote 14
Further still, Glass (Reference Glass2007: 89) now notes “the rightward stroke at the foot of k, g, ṭ, d, s is a defining feature […] rather than a foot mark”, a very welcome development in the understanding of the Kharoṣṭhī script, but the phonetic realization is still considered uncertain, saying it is “possible but not certain that this scribe used ḵ/g̱ to phonetically distinguish intervocalic k and g […] the evidence is inconclusive on the basis of this manuscript alone” (Glass Reference Glass2007: 107), and yet further down “these forms, and the developments of intervocalic g, suggest that the pair g and g̱ were minimally distinguished” (Glass Reference Glass2007: 119). In spite of Glass’s remarks regarding the stroke at the bottom of g̱ as a feature of the character and not a flourish, Lenz (Reference Lenz2010) still does not distinguish it and transcribes all instances as g, in spite of being published in the same collection with the same editorial team. In the most recent volume of the collection, Schlosser (Reference Schlosser2022: 81) alludes to this issue, saying “When Gāndhārī manuscripts were initially being studied, g and g̱ were not differentiated consistently by every editor because the two signs did not imply a difference in meaning. Since they do reflect a phonological difference, in this publication a distinction is maintained”.
In light of the fluctuation in the recent scholarship, and answering calls by Brough (Reference Brough1962: 85) and Glass (Reference Glass2007: 107) to conduct an analysis of the distribution of the spellings for intervocalic velar stops, we first look at the outcome of the most commonly found intervocalic k, namely the suffix -ka. To this end, we tally the attestations of Gāndhārī words corresponding to Sanskrit words ending in -ka, and we group these attestations by the vowel that precedes -ka and by text.
Since the two main vowels preceding -ka are a and i, we first present these two in Table 1. The texts have been ordered to highlight the pattern: on the -aka# side (left), we can see four distinct groups: from the top to CKI 48, the outcome is predominantly k; from NirdL to DhpSp, the outcome is predominantly g; from CKD 72 to CKD 14, g̱; finally, from AnavL to the bottom, the main outcome is an empty spelling (“∅”, such as uasao in SĀS6, from Skt. upāsaka “lay follower”). On the -ika# side (right), however, and with the exception of the two oldest texts (Shāh and Mān), there is a very clear predominance of either an empty spelling or a y.Footnote 15
Table 1. Distribution of the Gāndhārī outcomes for the k in Skt. -aka# (left) and -ika# (right)
Such a contrast between these two environments tells us several things:
• First, it cannot be regarded as correct that Gāndhārī scribes, on the whole, spelled randomly, sometimes transcribing the actual pronunciation (y /j/ or ∅), and sometimes somehow using g or g̱ to indicate knowledge of an original k that would now be pronounced /j/. This is not to say that historical spellings do not exist in Gāndhārī, they certainly do, but on the whole “historical spelling” should be our last recourse when trying to explain seemingly chaotic spelling behaviours in Gāndhārī.
• Second, and now considering that Gāndhārī spelling represents pronunciation with a higher fidelity than previously assumed, the Gāndhārī corpus can be segmented in at least four groups, based on the outcome of k, and studies of other phonological or linguistic features can evaluate their material against this segmentation.
• Finally, at least for the “k” and “g” segments of the corpus, the outcome of k was dependent on the preceding vowel and led to at least two different phonemes: preceded by i, k underwent palatalization and yielded /j/, represented by y or a gap, and that /j/ might itself have disappeared. The outcome of k preceded by a depends on the document, with the g and g̱ indicating a lenition of sorts, either through voicing to /g/ or fricativization (with or without voicing) to /x/ ∼/χ/ or /ɣ/.
Combined with previous rules, such as compound boundaries (including directional prefix boundaries, most of the time), this conditioning environment explains most of the data, sometimes even at the text level: Salomon (Reference Salomon2008: 108) comments that the outcome of k “seems to be more or less arbitrary and unpredictable; compare, for example, the two different reflexes in the etymologically cognate aṇomia = *anavamikāḥ and aṇom[a]gha = *anavamaka-”, providing a minimal pair where the difference in the k-preceding vowel entirely conditions the outcome. Similarly, although the data is not entirely clean, a look at NirdL2 in Table 1 suggests that Baums’ (Reference Baums2009) comment “[t]he distribution of the three spellings [g, y or ∅] does not follow any discernible pattern, and it has to be assumed that the same pronunciation is intended everywhere” (Baums Reference Baums2009: 140) needs revision.
An important question emerges: for the g and g̱ segments (and to a lesser extent for the k segment), what was the phonetic realization of the outcome of k preceded by a? As noted above, the general belief is that g̱ represented a fricative, and Glass (Reference Glass2007: 119) says the distribution of g and g̱ “suggest[s] that the pair g and g̱ were minimally distinguished”, which we can but agree with. Provisionally, we think adopting /ɣ/ as a realization of g̱ is most reasonable. What about the “g texts”? Because neither NirdL2 nor EĀBm spells g̱ in any context, it is difficult to say whether the scribes simply lacked the means to represent the phonetic outcome of k and went with what sounded closest, grouping these texts with the g̱ ones, or whether the outcome was actually /g/ and they should be considered a group of their own. The fact that nearly all of the texts in the g̱ group were found in Niya could just as well suggest a distinct dialect or simply a local spelling innovation not found in the g group, and a comparison of other phonological developments would be necessary to help decide this issue.
4.2.3. Intervocalic k in other environments
The contrast between final -aka and final -ika established, we now turn to other environments, both non-final environments and other k-preceding vowels.
In the Gāndhārī corpus, we could find no real example of non-final -aka or -ika: all word-internal cases fall into one of the previously studied cases, namely compound boundary (NirdL2 kriṣ̄akamo < Skt. kṛṣṇakarman “dark action”) and “word-final” in compounds. We do, however, find cases of k preceded by the other vowels e, o, and u.
It has been widely notedFootnote 16 that the word eka “one” is always spelled with a k in Gāndhārī, a phenomenon generally explained by saying that the word must have had a geminate velar – ekka – which would have prevented lenition. What is interesting and, to the best of our knowledge, has never been mentioned, is the fact that words based on this eka “one” such as Skt. pratyeka “one by one, separately” or Skt. aneka “many (lit. not one)” do lenite as one would expect, according to their group.Footnote 17
Beyond this note, the outcome of k when preceded by either e, o, or u is the same as when preceded by a, as shown in Table 2. We note that there is an unexplained tendency in the “y/∅” segment of the table to spell Skt. -ok- words (right side of the table) with a g or g̱ instead of the expected y or ∅, a tendency not seen in -ek- and -uk- (left side). It is perhaps interesting that the conditioning of the outcome of Skt. k lies on a line between i on one side and all other vowels on the other side.
Table 2. Distribution of the Gāndhārī outcomes for the k in Skt. -ek- and -uk- (left) and -ok- (right)
4.2.4. Intervocalic
With the outcome of Skt. k covered in intervocalic environments, a comparison of the outcome of Skt. g in similar environments is possible. As we find no “i vs all other vowels” contrast for g, the results are collapsed in Table 3.
Table 3. Distribution of the Gāndhārī outcomes for the g in Skt. -ag-, -eg-, -ig-, -og-, and -ug-
Unfortunately, the results are somewhat difficult to interpret. The general shape of Table 3 resembles the previous tables, with the obvious difference that documents that had Skt. k preserved now have Skt. g preserved. Beyond this expected difference, which groups the so-called “k and g segments” together, the g̱ segment still shows distinctively.Footnote 18 As for the so-called “y/∅” segment, the inconsistency deserves some comment: for MĀS, Silverlock (Reference Silverlock2015: 174) notes that Skt. bhagavant is spelled bhag̱ava in the body of the sutra, but bhayava in the standard closing formula, as is seen throughout the collection of manuscripts of the so-called Senior Scribe. He then proposes that the sutra was copied from an archetype that had the g̱ spelling – representing a by-then dated pronunciation – while closing formulae were fixed and “integral to the memorisation process of reciters”. This spelling of bhayava also explains the 30 occurrences of y in the SĀS in Table 3.
Overall, the main takeaway from Table 3 is that the outcome of Skt. g is not spelled the same as that of Skt. k:
• As opposed to Skt. -ik-, Skt. -ig- does not commonly yield y/∅.
• As opposed to Skt. -ak-, -ek-, -ok-, and -uk-, Skt. -ag-, -eg-, -og-, and -ug- do not regularly yield y/∅ in the “y/∅ texts”, and the distribution of attestations across a text – such as the distinction between the sutra and the closing formula mentioned above for the corpus of the Senior scribe (SĀS and MĀS) – is not sufficient to explain the distinction between the outcome of k and g preceded by these vowels.
This precludes k > g/V _ V, since otherwise the outcome of k and g would be identical. Instead, k must have fricativized first, with k > j/i _ V happening very early: this is the stage witnessed in the so-called “k documents”. This was then followed by k > ɣ/V _ V (where the first V could not be /i/ any more). Then, some dialects must have undergone ɣ > j/V _ V while others did not, leading to the distinction between the “g and g̱ texts” on the one hand, and the “y/∅ texts” on the other, after both had undergone g > ɣ/V _ V.
5. Conclusion
In contrast with editions of individual texts, this study has approached the question of the development of Sanskrit velar stops in Gāndhārī from a corpus-wide angle. In doing so, we were able to highlight trends in the corpus and better understand the possible development of Gāndhārī phonology.
The development of velar stops is summarized below. While it is not possible to establish a relative chronology between every pair of sound changes listed below, we indicate chronological constraints when known.
• voiceless aspirated kh:
kʰ > h / V _ V
kʰ > gʱ / [ +nasal ] _
skʰ > kʰ
• voiced aspirated gh:
gʱ > h / V _ V
\begin{align*}
{\textrm{g}}^\text{h} \gt \left\{\begin{array}{l}
{\textrm{g}}^\text{h}/{\#}{\_{\textrm{r}}} \\
{\textrm{g}}/{\#}{\_}
\end{array}\right.
\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\ \ \end{align*}
• voiceless unaspirated k:
k > j / i _ V (in all but the very earliest texts)
k > ɣ > j / V _ V (with texts showing the three stages)
k > g / [ +nasal ] _
sk > ʰk > kʰ
k > ∅ / _ #
kt > tː
-
\begin{equation*}
\text{k}\upsilon \gt \left\{\begin{array}{l}
\text{k}\upsilon/{\#}{\_}\\
\text{K}
\end{array}\right.\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\quad\quad\ \ \end{equation*}
• voiced unaspirated g:
g > g / # _
g > ɣ > j / V _ V (only some texts show the second development; this chain visibly started after the corresponding k > ɣ > j / V _ V had completed)
On a historical phonology level, the main contribution of the article is the demonstration that the Sanskrit unaspirated stops k and g followed separate development chains in intervocalic contexts, and that in particular k followed a separate development depending on whether the preceding vowel was i or not (while the development of g shows no such conditioning). Such a discovery sheds light on why some Gāndhārī loanwords into Chinese are transcribed with characters with a final -k while others are not: Skt. campaka “of Campā” > Gdh. *caṃpag̱(a) > Late Hàn Chinese 占匐 */tśam.bǝk/ with a final /-k/,Footnote 19 while Skt. sugandhika “some type water lily” > Gdh. *sugaṃdhi(a) > LHC 須揵提 */sio.gan.de/ without such a coda.
Beyond this specific phonological discovery, the main contribution of the article is to demonstrate that, overall, Gāndhārī scribes spelled with a much higher fidelity to pronunciation than is usually considered, and that attempting to discard “historical spelling” as a primary explanation is a promising approach that will yield new insights in Gāndhārī linguistics.
Acknowledgements
We would like to extend our thanks to Stefan Baums and Andrew Glass for permission to reproduce Figures 1 and 2, and the British Library for Figures 4, 5, 6, 7, 8, 9, 10, 11, and 12. For Figure 3, the original manuscript seems to have disappeared and its current location is unknown.
Funding
This work was funded by the Arts and Humanities Research Council (AHRC), UKRI, as part of the project “Han Phonology: When Chinese Became Chinese”. Project Reference: AH/V008722/1. Principal Investigator: Nathan Hill, SOAS University of London, Trinity College Dublin.
Appendix A. Date and findspot of the documents used in the article
Table 4 provides a reference to the dates and findspots attributed to the documents found in Table 1, 2, and 3 of the article, retrieved from Baums and Glass (2002).Footnote 20 The information can be used to study chronological and geographical trends in the development of the intervocalic unaspirated velars.
Appendix B. Proposed corrections to text transcriptions
B1. MSŪB
Besides the seven attestations of bhag̱ava for Skt. bhagavant, Strauch (Reference Strauch2010: 28) transcribes the attestation of line 9 as bhagava (with a plain g). As shown in Figure 3, this is a mistake, and bhag̱ava should be read on line 9 like the two occurrences on line 12 (and tasag̱adaśaśa- on line 10). This means the scribe was consistent in the spelling of this word.
Figure 3. Lines 9–12 of the MSūB manuscript, as found in Strauch (Reference Strauch2010: 28)
B2. CKD 140
One instance of Gdh. aroga (< Skt. aroga “healthy”) is found in CKD 140 and transcribed in Boyer et al. (Reference Boyer, Rapson and Senart1920: 56) as arogemi. Based on the manuscript shown in Figure 4, and comparing with the other two instances of g̱ (then transcribed ǵ), we choose to read arog̱emi, although that reading is less certain than that of bhag̱ava in Appendix B.1.
Figure 4. Lines 1–6 of the under-tablet obverse of CKD 140, as found in Stein (Reference Stein1907)
B3. AVL6
Lenz (Reference Lenz2003: 118) – commenting on the spelling of g with a foot mark and on its usual attribution of the sound /ɣ/ – writes “a distinct phoneme is probably not indicated in our text”. As a consequence, and in contrast with the other volumes of the Gandhāran Buddhist Texts collection, Lenz (Reference Lenz2003) does not transcribe g̱ separately from g. We would like to show that the spelling of this text is in fact very consistent with the phonological development of velar stops presented in this article.
We find the ga akṣara written with six different shapes (shown in Figure 5): with a straight vertical stem, with a left “tail”, with a right “tail”, with a stroke from left to right of the vertical stem, with a right curl continuing the vertical stem, and with a clockwise loop crossing the stem. Our analysis does not differ from the generally accepted palaeographic analysis of ga and its variations:
Figure 5. The various shapes of the ga akṣara: (a) plain, (b) left tail, (c) right tail, (d) left-to-right stroke, (e) curl, (f) rga
• The left tail is merely a flourish and represents the same sound as a plain vertical stem, as seen in Figure 6 in two words having g in word-initial position: ga[do] (plain stem, Skt. gata “gone”) and gilanag̱o (left tail, Skt. glāna+ka “sick”).
• The right tail and left-to-right stroke shapes represent /ɣ/, which we transcribe as g̱, as seen in Figure 7 in two words with expected /ɣ/: nag̱are (right tail, Skt. nagara “city”) and likhidag̱e (left-to-right stroke, Skt. likhita+ka “written”).
• In this text, the shape with a right curl represents gr, as seen in Figure 8: anugrahartho (Skt. anugrahārtham “as a favour”).
• The shape with a clockwise loop represents rga, see for instance Glass (Reference Glass2000, 62).
Figure 6. Two words with an initial /g/: (a) ga[do], (b) gilanag̱o
Figure 7. Two words with an expected /ɣ/: (a) nag̱are, (b) likhidag̱e
Figure 8. Word with an expected gr: anugrahartho
On this basis, we analyse all instances of g in the transcription of the text, and propose a revised transcription of the words. When the reading is uncertain, we transcribe according to the expected outcome and indicate our uncertainty with square brackets. The original transcription contains 54 instances of g across 51 words, listed in Table 5. For reference, the manuscript found in Lenz (Reference Lenz2003) is reproduced here, with the relevant words underlined, in Figures 9, 10, 11, and 12.
Figure 9. AvL6, lines 1–19
Figure 10. AvL6, lines 19–34
Figure 11. AvL6, lines 35–48
Figure 12. AvL6, lines 47–61
Table 5. Revised transcription of words originally transcribed with g in Lenz (Reference Lenz2003)
The scribe’s spelling shows a few peculiarities:
• Sanskrit words that have g or k preceded by a long ā, like Skt. hastināga “stately elephant”, are spelled with a plain g with complete regularity (five occurrences). This seems to be a conditioning factor blocking further lenition, and this contrasts with the rest of the corpus where the length of the preceding vowel seems to make no difference to the outcome of the velar stop.
• The outcome of Skt. eka “one” is eg[o], and not eka, as is found elsewhere in most of the Gāndhārī corpus.
• The name of the Saka people is spelled as sago. This form is probably not inherited from Sanskrit (Skt. Śaka) given the absence of ś, and the close proximity of the Saka people with Gāndhārī speakers may have led to a direct borrowing at a later date than the normal development of k > ɣ/a _ V.
• Except in rayo and rayam (< Skt. rājā “king”), expected y as the initial of the last syllable is never spelled y. Instead, the two instances of a locative (L27 and L28), the words in Skt. -iya and -īya are spelled with a g, while the words in Skt. -ika, -ija, and -aya are spelled with g̱. Since the scribe’s spelling – as we will discuss below – is otherwise completely consistent with pronunciation, this spelling practice of rendering all expected y as g and g̱ suggests a possible fortition of y /j/ to a palatal fricative, phonetically close to the realization of g̱.
These peculiarities aside, the scribe’s spelling is extremely consistent with the expected spelling of a “g̱ text”:
• Out of 22 instances of expected g̱,Footnote 21 we find 19 instances of g̱ and three cases where the manuscript is damaged right at the foot of the akṣara or the ink has faded too much; which translates to a complete consistency of the spelling for the cases where g̱ would be expected.
• Of the nine instances of expected g, seven do indeed show g. The exceptions are istrig̱ag̱anapradivrud[o] mentioned above, and agachadi (Skt. āgacchati “(he) comes”), for which the reading is uncertain, resulting in seven out of eight cases where we have enough confidence.
This conclusion, that AvL6 is a “g̱ text”, reduces the already small body of texts belonging to the “g texts”, and suggests we need to re-visit the other transcriptions produced by Lenz following the same transcription conventions (Lenz Reference Lenz2010, Reference Lenz and Collett2014).
Open access
